Diagnostic performance of a deep learning convolutional neural network in the differentiation of combined naevi and
Background:
Deep learning convolutional neural networks (CNN) may assist physicians in the diagnosis of melanoma. The capacity of a CNN to differentiate melanomas from combined naevi, the latter representing well-known melanoma simulators, has not been investigated.
Objective:
To assess the diagnostic performance of a CNN when used to differentiate melanomas from combined naevi in comparison with dermatologists.
Methods:
In this study, a CNN with regulatory approval for the European market (Moleanalyzer-Pro, FotoFinder Systems GmbH, Bad Birnbach, Germany) was used. We attained a dichotomous classification (benign, malignant) in dermoscopic images of 36 combined naevi and 36 melanomas with a mean Breslow thickness of 1.3 mm. Primary outcome measures were the CNN's sensitivity, specificity and the diagnostic odds ratio (DOR) in comparison with 11 dermatologists with different levels of experience.
Results:
The CNN revealed a sensitivity, specificity and DOR of 97.1% (95% CI [82.7-99.6]), 78.8% (95% CI [62.8-89.1.3]) and 34 (95% CI [4.8-239]), respectively. Dermatologists showed a lower mean sensitivity, specificity and DOR of 90.6% (95% CI [84.1-94.7]; P = 0.092), 71.0% (95% CI [62.6-78.1]; P = 0.256) and 24 (95% CI [11.6-48.4]; P = 0.1114). Under the assumption that dermatologists use the CNN to verify their (initial) melanoma diagnosis, dermatologists achieve an increased specificity of 90.3% (95% CI [79.8-95.6]) at an almost unchanged sensitivity. The largest benefit was observed in 'beginners', who performed worst without CNN verification (DOR = 12) but best with CNN verification (DOR = 98).
Conclusion:
The tested CNN more accurately classified combined naevi and melanomas in comparison with trained dermatologists. Their diagnostic performance could be improved if the CNN was used to confirm/overrule an initial melanoma diagnosis. Application of a CNN may therefore be of benefit to clinicians.
